CUDA函数不会在带有Numba的Python上执行For循环 [英] CUDA Function Won't Execute For Loop on Python with Numba
问题描述
我正在尝试在GPU上运行模拟的简单更新循环。基本上有很多由圆圈表示的生物，它们在每个更新循环中都会移动，然后将检查它们是否相交。
I'm trying to run a simple update loop of a simulation on the GPU. Basically there are a bunch of "creatures" represented by circles that in each update loop will move and then there will be a check of whether any of them intersect.
import numpy as np
import math
from numba import cuda
@cuda.jit('void(float32[:], float32[:], float32[:], uint8[:], float32[:], float32[:], float32, uint32, uint32)')
def update(p_x, p_y, radii, types, velocities, max_velocities, acceleration, num_creatures, cycles):
for c in range(cycles):
for i in range(num_creatures):
velocities[i] = velocities[i] + acceleration
if velocities[i] > max_velocities[i]:
velocities[i] = max_velocities[i]
p_x[i] = p_x[i] + (math.cos(1.0) * velocities[i])
p_y[i] = p_y[i] + (math.sin(1.0) * velocities[i])
for i in range(num_creatures):
for j in range(i, num_creatures):
delta_x = p_x[j] - p_x[i]
delta_y = p_y[j] - p_y[i]
distance_squared = (delta_x * delta_x) + (delta_y * delta_y)
sum_of_radii = radii[types[i]] + radii[types[i]]
if distance_squared < sum_of_radii * sum_of_radii:
pass
acceleration = .1
creature_radius = 10
spacing = 20
food_radius = 3
max_num_creatures = 1500
num_creatures = 0
max_num_food = 500
num_food = 0
max_num_entities = max_num_creatures + max_num_food
num_entities = 0
cycles = 1
p_x = np.empty((max_num_entities, 1), dtype=np.float32)
p_y = np.empty((max_num_entities, 1), dtype=np.float32)
radii = np.array([creature_radius, creature_radius, food_radius], dtype=np.float32)
types = np.empty((max_num_entities, 1), dtype=np.uint8)
velocities = np.empty((max_num_creatures, 1), dtype=np.float32)
max_velocities = np.empty((max_num_creatures, 1), dtype=np.float32)
# types:
# male - 0
# female - 1
# food - 2
for x in range(1, 800 // spacing):
for y in range(1, 600 // spacing):
if num_creatures % 2 == 0:
types[num_creatures] = 0
else:
types[num_creatures] = 1
p_x[num_creatures] = x * spacing
p_y[num_creatures] = y * spacing
max_velocities[num_creatures] = 5
num_creatures += 1
device_p_x = cuda.to_device(p_x)
device_p_y = cuda.to_device(p_y)
device_radii = cuda.to_device(radii)
device_types = cuda.to_device(types)
device_velocities = cuda.to_device(velocities)
device_max_velocities = cuda.to_device(max_velocities)
update(device_p_x, device_p_y, device_radii, device_types, device_velocities, device_max_velocities,
acceleration, num_creatures, cycles)
print(device_p_x.copy_to_host()[0])
math.cos和math.sin中的1.0只是单个生物
方向的占位符我有一个围绕循环执行的循环次数。如果我尝试删除它，而只留下代码块来移动这些生物，即使我向它们添加常数，p_x，p_y或速度都不会改变。为什么不呢？
The 1.0 in math.cos and math.sin is just a placeholder for the directions of the individual creatures I have a surrounding loop executed cycles amount of times. If I try to remove it and only leave the block of code moving the creatures neither p_x, p_y or velocities have changed, even if I add a constant to them. Why not?
推荐答案
至少有两个问题：
-
您不是要初始化
速度：
velocities = np.empty((max_num_creatures, 1), dtype=np.float32)
我们可以通过以下方法来对其进行简单的测试：
we can fix that for a trivial test with:
velocities = np.ones((max_num_creatures, 1), dtype=np.float32)
这不是正确的数组形状：
This isn't the correct array shape:
p_x = np.empty((max_num_entities, 1), dtype=np.float32)
^^^^^^^^^^^^^^^^^^^^^
匹配您的内核签名：
@cuda.jit('void(float32[:], float32[:], float32[:], uint8[:], float32[:], float32[:], float32, uint32, uint32)')
^^^^^^^^^^
我们可以通过以下方式解决此问题：
we can fix that with:
p_x = np.empty(max_num_entities, dtype=np.float32)
，对于 p_y ，类型也是如此，速度和 max_velocities 。 （我想可能还会对 radio 进行一些更改，但是并不清楚您打算做什么，因为看起来您想要一个多维数组，但是正在以一维数组AFAICT的形式在内核中进行访问。此外，内核代码的这一部分是无能为力的，因此与手头的问题或多或少无关紧要。
and likewise for p_y, types, velocities, and max_velocities. (I imagine some change may possibly be in order also for radii, but it's not entirely clear what you intend with that, since it appears you want a multi-dimensional array, but are accessing it in-kernel as a single-dimensional array, AFAICT. Furthermore, that section of your kernel code is a do-nothing, so it is more or less irrelevant for the problem at hand).
当我进行这些更改时，我得到的似乎是合理的输出：
When I make those changes, I get what appears to be rational output:
$ cat t9.py
import numpy as np
import math
from numba import cuda
@cuda.jit('void(float32[:], float32[:], float32[:], uint8[:], float32[:], float32[:], float32, uint32, uint32)')
def update(p_x, p_y, radii, types, velocities, max_velocities, acceleration, num_creatures, cycles):
for c in range(cycles):
for i in range(num_creatures):
velocities[i] = velocities[i] + acceleration
if velocities[i] > max_velocities[i]:
velocities[i] = max_velocities[i]
p_x[i] = p_x[i] + (math.cos(1.0) * velocities[i])
p_y[i] = p_y[i] + (math.sin(1.0) * velocities[i])
for i in range(num_creatures):
for j in range(i, num_creatures):
delta_x = p_x[j] - p_x[i]
delta_y = p_y[j] - p_y[i]
distance_squared = (delta_x * delta_x) + (delta_y * delta_y)
sum_of_radii = radii[types[i]] + radii[types[i]]
if distance_squared < sum_of_radii * sum_of_radii:
pass
acceleration = .1
creature_radius = 10
spacing = 20
food_radius = 3
max_num_creatures = 1500
num_creatures = 0
max_num_food = 500
num_food = 0
max_num_entities = max_num_creatures + max_num_food
num_entities = 0
cycles = 1
p_x = np.empty(max_num_entities, dtype=np.float32)
p_y = np.empty(max_num_entities, dtype=np.float32)
radii = np.array([creature_radius, creature_radius, food_radius], dtype=np.float32)
types = np.empty(max_num_entities, dtype=np.uint8)
velocities = np.ones(max_num_creatures, dtype=np.float32)
max_velocities = np.empty(max_num_creatures, dtype=np.float32)
# types:
# male - 0
# female - 1
# food - 2
for x in range(1, 800 // spacing):
for y in range(1, 600 // spacing):
if num_creatures % 2 == 0:
types[num_creatures] = 0
else:
types[num_creatures] = 1
p_x[num_creatures] = x * spacing
p_y[num_creatures] = y * spacing
max_velocities[num_creatures] = 5
num_creatures += 1
device_p_x = cuda.to_device(p_x)
device_p_y = cuda.to_device(p_y)
device_radii = cuda.to_device(radii)
device_types = cuda.to_device(types)
device_velocities = cuda.to_device(velocities)
device_max_velocities = cuda.to_device(max_velocities)
update(device_p_x, device_p_y, device_radii, device_types, device_velocities, device_max_velocities,
acceleration, num_creatures, cycles)
print(device_p_x.copy_to_host())
$ python t9.py
[ 2.05943317e+01 2.05943317e+01 2.05943317e+01 ..., 3.64769361e-11
1.52645868e-19 1.80563260e+28]
$
还要注意，当前您仅启动一个线程的一个块，但我认为这与您的重新启动无关任务，目前。
Also note that currently you are only launching one block of one thread, but I assume that is not pertinent to your request, currently.
这篇关于CUDA函数不会在带有Numba的Python上执行For循环的文章就介绍到这了，希望我们推荐的答案对大家有所帮助，也希望大家多多支持IT屋！
